How well does an AI explain what it sees?
An AI system recognizes bird species in photos that have been damaged by noise, blur, fog, compression and similar effects. It explains each decision by pointing to regions of the photo and showing similar-looking regions from birds it saw during training.
You will see 16 short screens. Each shows one damaged bird photo and two different explanations (A and B) of the AI's decision. You pick which explanation is better (or "No clear difference" if you really cannot tell them apart). There are no right or wrong answers, and you do not need to know anything about birds.
- Takes about 8–10 minutes. Please use a laptop or desktop screen.
- Participation is voluntary and anonymous. We do not collect your name, email or any identifying information.
- You can stop at any time by closing the page; nothing is sent until you press Submit at the end.
- Your answers are used only for research on explainable AI and may be published in aggregate.
Your window looks narrow. The images are easier to compare on a wider screen.
Two quick questions about you
How to read an explanation
Each explanation is one row like the one below.
- Left: an undamaged copy of the same photo, with small coloured boxes marking the regions the AI relied on when it looked at the damaged photo. (We draw the boxes on the undamaged copy so you can see which parts of the bird they cover.) Boxes of one colour (and number) belong together.
- Right: five training examples. Example k has a frame of the same colour as group k on the left, and its own boxes show the region the AI thinks looks alike.
A good explanation puts its boxes on meaningful parts of the bird (head, beak, eye, wing, tail, body markings) and matches them to similar parts in the training examples. A poor one points at background, noise or damage, or matches regions that do not look alike.
Screen
Below are two explanations (A and B) of the AI's decision for this photo. Each one draws the AI's boxes on an undamaged copy of the photo. Compare where the boxes fall on the bird and how well they match the training examples, then answer the question at the bottom.
A good explanation puts its boxes on meaningful parts of the bird (head, beak, eye, wing, tail, body markings) and matches them to similar parts in the training examples. A poor one points at background, noise or damage, or matches regions that do not look alike.
Tip: boxes and frames with the same colour and number belong together.
Almost done
Thank you!
We could not reach our server. Please copy the code below and send it to the person who shared this study with you.